Falcon

This model was released on 2023-11-28 and added to Hugging Face Transformers on 2023-07-11.

Falcon

Falcon is a family of large language models, available in 7B, 40B, and 180B parameters, as pretrained and instruction tuned variants. This model focuses on scaling pretraining over three categories, performance, data, and hardware. Falcon uses multigroup attention to significantly reduce inference memory requirements and rotary positional embeddings (RoPE). These models are pretrained on RefinedWeb, a high-quality and deduplicated 5T token dataset.

You can find all the original Falcon checkpoints under the Falcon collection.

Tip

Click on the Falcon models in the right sidebar for more examples of how to apply Falcon to different language tasks.

The example below demonstrates how to generate text with Pipeline, AutoModel, and from the command line.

import torch
from transformers import pipeline

pipeline = pipeline(
    task="text-generation",
    model="tiiuae/falcon-7b-instruct",
    dtype=torch.bfloat16,
    device=0
)
pipeline(
    "Write a short poem about coding",
    max_length=100,
    do_sample=True,
    temperature=0.7
)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-7b-instruct")
model = AutoModelForCausalLM.from_pretrained(
    "tiiuae/falcon-7b-instruct",
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
)

input_ids = tokenizer("Write a short poem about coding", return_tensors="pt").to(model.device)

output = model.generate(**input_ids)
print(tokenizer.decode(output[0], skip_special_tokens=True))

# pip install -U flash-attn --no-build-isolation
transformers chat tiiuae/falcon-7b-instruct --dtype auto --attn_implementation flash_attention_2 --device 0

Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the Quantization overview for more available quantization backends.

The example below uses bitsandbytes to only quantize the weights to 4-bits.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-7b")
model = AutoModelForCausalLM.from_pretrained(
    "tiiuae/falcon-7b",
    dtype=torch.bfloat16,
    device_map="auto",
    quantization_config=quantization_config,
)

inputs = tokenizer("In quantum physics, entanglement means", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Notes

• If you’re upgrading from an older custom code checkpoint, remember to convert it to the official Transformers format for better stability and performance using the conversion script located in the Falcon model directory.

  python convert_custom_code_checkpoint.py --checkpoint_dir my_model

FalconConfig

[[autodoc]] FalconConfig - all

FalconModel

[[autodoc]] FalconModel - forward

FalconForCausalLM

[[autodoc]] FalconForCausalLM - forward

FalconForSequenceClassification

[[autodoc]] FalconForSequenceClassification - forward

FalconForTokenClassification

[[autodoc]] FalconForTokenClassification - forward

FalconForQuestionAnswering

[[autodoc]] FalconForQuestionAnswering - forward